Synthetic market research uses AI-generated personas to simulate how audiences respond to surveys, messages, campaigns, or strategic decisions, as an alternative or complement to recruiting human participants. It is best understood not as a wholesale replacement for traditional market research, but as a different tool with different strengths: traditional research remains superior for some questions, while synthetic research opens up others that were previously impossible. The field also spans several distinct methodological approaches that vary widely in how they build personas and how accurate the results are — a distinction set out in detail in Artificial Societies' method and evaluation.
Traditional market research is irreplaceable for a whole class of questions, and synthetic research is not designed to replace it. Anything that depends on physical, sensory, or hands-on experience — product packaging tests, usability and user-experience testing, taste and product trials, and in-person ethnography — requires real people interacting with real artefacts. The same holds wherever direct human interaction adds irreplaceable nuance, or where the goal is to observe genuine, unprompted behaviour. For understanding how everyday consumers experience a tangible product, a traditional panel or lab study remains the right and most cost-effective method. Synthetic research complements these approaches rather than competing with them.
Synthetic research is strongest precisely where traditional methods struggle. It can reach high-value audiences that are difficult or impossible to recruit — policymakers, executives, regulators, and other elites — and it can do so confidentially, testing sensitive strategies before anything enters the field. Its defining strength is opinion-shift modelling: establishing a baseline, exposing an audience to a message or decision, and measuring how opinion moves and spreads through a group. That capability makes possible a set of solutions traditional research cannot practically deliver — strategic communications, crisis and reputation, government affairs, investor relations, and innovation advantage — where the value lies in anticipating reactions to consequential decisions before they are made. And because it is simulation, organisations can test the full range of options, not only the single candidate they could afford to put into the field.
Synthetic research is not a single methodology, and the approaches differ sharply in rigour. The simplest prompt a large language model with an invented biography ("you are a 34-year-old teacher from Ohio"), which tends to produce generic, stereotyped answers. Others build top-down synthetic populations that reproduce headline averages but invent the individuals beneath them, or bottom-up digital twins that capture individual nuance but miss population-level diversity. Artificial Societies takes a different approach again — networks of enriched personas grounded in real-world observations and connected by network science — designed to capture accurate top-line distributions and rich individual nuance at the same time. These trade-offs, and the accuracy each approach achieves, are compared in full in the method and evaluation.
Artificial Societies constructs networks of 200 to 3,500 interconnected AI personas that represent a specific real-world audience. Each persona is built from anonymised real-world observations — what real people say, do, and engage with — rather than demographic stereotypes, and the personas are connected in a social graph so the model captures how opinion forms and spreads, not just what individuals think in isolation. The result is research that opens all the way down to segments, drivers, and individual voices, so a recommendation can be interrogated and defended. Full detail on how the personas are built, and how the approach performs against human panels, is available in the method and evaluation.
Synthetic market research uses AI-generated personas to simulate how audiences would respond to surveys, messages, or strategic decisions, instead of — or alongside — recruiting human participants. It is most valuable where traditional research cannot practically reach: hard-to-reach audiences, confidential materials, tight timelines, and modelling how opinion shifts in response to a decision.
No. The two are complementary. Traditional research remains superior for physical and experiential questions — product packaging, usability testing, taste trials, in-person ethnography — while synthetic research is stronger for reaching hard-to-reach audiences, testing confidential strategies, and modelling how opinion shifts. The right method depends on the question.
They range from biography-prompted LLMs (an invented biography fed to a language model, which produces generic answers), to top-down synthetic populations (accurate averages but invented individuals), to bottom-up digital twins (individual nuance but limited population diversity), to Artificial Societies' networks of enriched personas (accurate distributions and individual nuance together). The method and evaluation sets out how each performs.
It is best suited to consequential decisions where getting audience response wrong is costly and traditional research is too slow, too expensive, or too risky — such as strategic communications, crisis and reputation, government affairs, investor relations, and innovation advantage.